harbor-framework/harbor

Harbor solves the fragmentation and reproducibility problem in agent evaluation by providing a unified harness that works with diverse agents (e.g., Claude Code, OpenHands) and benchmarks (e.g., Terminal-Bench, SWE-Bench), while supporting parallel execution and RL rollout generation.

Developers can evaluate any agent against existing benchmarks, create and share custom benchmarks, run thousands of parallel environments using providers like Daytona and Modal, and generate rollouts for reinforcement learning optimization. Real-world use cases include benchmarking coding agents, testing model improvements, and scaling agent evaluation for research.